@verdicter/sdk

Runtime security for AI agents. Evaluate every tool call against your policies - get ALLOW, DENY, or ESCALATE in under 50ms.

What is Verdicter?

Your agents act fast. Verdicter makes sure they act right.

Every tool call - across every agent, every request - passes through a policy engine in real time before it executes. No per-call code changes needed.

Agent action → verdicter.evaluate() → Policy engine → ALLOW / DENY / ESCALATE → Tool runs

→ Get your free API key at verdicter.dev

Install

npm install @verdicter/sdk

Quick start

1. Get an API key at verdicter.dev - free to sign up, 10k evaluations/month on the free plan.

2. Register your agent in the dashboard and create a policy (e.g. "block any send_email where recipient is external").

3. Evaluate before every tool call:

import { Verdicter } from '@verdicter/sdk';

const verdicter = new Verdicter({
  apiKey: process.env.VERDICTER_API_KEY!,
});

const { decision, modifiedPayload } = await verdicter.evaluate({
  agentId: 'support_bot',         // registered in your Verdicter dashboard
  tool:    'send_email',
  payload: { to: user.email, subject, body },
});

if (decision === 'ALLOW')    await sendEmail(payload);
if (decision === 'DENY')     throw new Error('Blocked by policy');
if (decision === 'ESCALATE') await requestHumanApproval(payload);

Or wrap your tools - zero per-call changes:

const safeSendEmail = verdicter.wrapFn('send_email', sendEmail, {
  agentId: 'support_bot',
});

// Evaluation + policy enforcement happens automatically
await safeSendEmail({ to: user.email, subject, body });

LangChain adapter

import { VerdicterToolkit } from '@verdicter/sdk/langchain';

const toolkit = new VerdicterToolkit({ client: verdicter, agentId: 'support_bot' });
const guardedTools = toolkit.guardTools(tools); // wrap your existing LangChain tools

Vercel AI SDK adapter

import { guardTools } from '@verdicter/sdk/vercel-ai';

const tools = guardTools(verdicter, 'support_bot', {
  send_email: tool({ ... }),
});

Configuration

const verdicter = new Verdicter({
  apiKey:      process.env.VERDICTER_API_KEY!,
  timeout:     5000,    // ms, default 5000
  maxRetries:  2,       // default 2
  failOpen:    false,   // if true, ALLOW on network errors (default: false = fail closed)
});

| Option | Type | Default | Description | |--------|------|---------|-------------| | apiKey | string | required | Your Verdicter API key | | timeout | number | 5000 | Request timeout in ms | | maxRetries | number | 2 | Retries on 429/5xx | | failOpen | boolean | false | ALLOW on network errors instead of throwing |

Decisions

| Decision | Meaning | |----------|---------| | ALLOW | Policy passed - run the tool | | DENY | Policy blocked it - don't run, inform user | | ESCALATE | Needs human review - route to your approval flow |

Dashboard

Everything is visible in your Verdicter dashboard:

• Live audit log - every evaluation with risk score, decision, and trace
• Policy editor - create rules in plain language or JSON
• Agents - register agents, track risk scores over time
• Escalations - approve/reject high-risk actions from Slack, email, or the dashboard
• Comply - SOC 2, GDPR, HIPAA reports generated from your audit log

Links

• Sign up free → verdicter.dev
• Dashboard → app.verdicter.dev
• Documentation → docs.verdicter.dev
• npm → npmjs.com/package/@verdicter/sdk

MIT License © Verdicter